VIDEO ANOMALOUS BEHAVIOUR DETECTION BASED ON COMPRESSED-INFLATED ATTENTION MODULE
DengBin Xu, PeiChen Wu, LiNing Yuan · Journal of Computer Science and Electrical Engineering · 2024
The performance of current attention based feature fusion methods depends on the correlation between features. After feature fusion, due to the inter domain differences of different features, the spatiotemporal perception ability is insufficient, and effectively fusing two cross domain features still faces challenges. A video anomaly detection method based on compression inflation attention feature fusion is proposed to address the issues of insufficient cross domain expression ability of RGB features and optical flow features, as well as weak spatiotemporal perception ability of fused features. The use of Squeeze and Inflation Networks (SENet) to construct a fusion mechanism for RGB and optical flow features can enhance the expression ability of fused features while reducing the number of network parameters and improving the performance of anomaly detection algorithms. In the global spatiotemporal awareness stage, the ConvLSTM (Long Short Term Memory Convolutional Network) is used to achieve global spatiotemporal awareness, while balancing computational complexity and detection performance. We achieved a recognition performance of 93.72% on the UCSDPed2 dataset and also performed well on the CUHKAvenue and LAD2000 datasets, verifying the effectiveness of the method.